Graph-based recommendation systems have excelled in modeling user–item interactions, but most focus solely on positive feedback, overlooking the critical role of negative feedback in capturing comprehensive user preferences. Signed graphs, which incorporate both positive and negative interactions, offer a promising approach but face challenges due to the oversimplification of balance theory and the limitations of conventional graph neural networks (GNNs) in processing negative signals. To address these issues, we propose paired channel enhanced sign-aware graph recommendation (PCSRec), a novel framework that holistically integrates positive and negative feedback. PCSRec introduces a path-enhanced embedding module that leverages a learnable path-encoding matrix to capture indirect structural patterns, overcoming the limitations of balance theory. Additionally, it employs a paired channel filtering mechanism with spectral low-pass and high-pass filters to model similarity from positive feedback and dissimilarity from negative feedback, respectively. A dual-loss optimization strategy, combining contrastive and Bayesian personalized ranking (BPR) losses, further refines discriminative representations. Extensive experiments on four real-world datasets demonstrate that PCSRec outperforms both unsigned and signed graph-based baselines, achieving state-of-the-art recommendation performance. Ablation studies and visualizations confirm the effectiveness of its components in improving embedding quality and recommendation accuracy.

In industrial applications, large models have exhibited superior generalization capabilities that are unattainable with smaller models. However, when faced with edge scenarios and highly diverse industrial samples, their deployment remains challenging due to high computational costs and unreliable output. To address these challenges, we propose CoLLM, a fuzzy large-small model collaborative framework, which dynamically selects between small and large models based on the characteristics exhibited by the samples. Specifically, this approach estimates uncertainty from input samples to guide model selection: low-uncertainty samples are processed by the small model for efficiency, while high-uncertainty or complex samples are routed to the large model for improved accuracy. It first constructs a fuzzy decision-making agent based on the fuzzy neural network (FNN) to assess sample complexity and determine the appropriate model for inference. Furthermore, a self-reflection mechanism is proposed to refine the large model's output, reducing the risk of unreliable output. Experimental results in industrial time series datasets demonstrate that our framework improves the computational efficiency of large models up to 14.54x while maintaining or improving prediction accuracy.

3-D modeling technologies play a crucial role in modern manufacturing. 3-D models are often exported as boundary representations for compatibility and data protection, which remove the modeling history and limit editability. To restore modeling sequences from such models, researchers employ neural networks to infer the possible modeling steps. This approach needs a large amount of labeled sequence data, and annotating such data is time-consuming. To address this issue, we propose a self-supervised pretraining method that generates modeling sequences directly from boundary representation models. Training data are first generated using a heuristic modeling sequences generation algorithm. Before training, each B-rep model is preprocessed into a zone graph representation. We then introduce the modeling operation evaluation network, which extracts features and scores each candidate operation to sequentially reconstruct the model. By selecting the most suitable operation at each step, the network progressively reconstructs the modeling sequence. This approach effectively reconstructs modeling sequences, restores the editability of B-rep models, and significantly reduces the reliance on labeled data.

Accurate and fast industrial time-series prediction is essential for safe and reliable operation of industrial equipment. Recent deep learning methods enable extracting complex temporal patterns by utilizing large-scale parameters and multiresolution feature extraction. However, they cause substantial computational complexity and limit their application at the edge. In this article, we design an adaptive temporal-channel multiresolution network (AMR-Net) that dynamically adjusts time-series resolution to avoid redundant computation. The motivation is that low-resolution feature representations are sufficient for predicting “easy” samples, whereas “hard” samples require high-resolution features to capture fine-grained information. For the AMR-Net, time series are initially input through a temporal-channel resolution decomposition (TCRD) module, which efficiently extracts low-resolution representations. Samples exhibiting high prediction confidence are expedited through early exit mechanisms, avoiding further processing. Meanwhile, high-resolution subnetworks capture the fine-grained information to discern the “hard” samples. Experiments on CMAPSS and N-CMAPSS datasets demonstrate that AMR-Net can improve computational speed by 15x while maintaining high accuracy.

With the continuous expansion of industrial systems, multisource and heterogeneous industrial data have increased rapidly, making the construction of a structured industrial knowledge system a core requirement in the industrial domain. Industrial knowledge graph (IKG) serves as a key approach for knowledge structuring and relation modeling and has become an indispensable foundation for industrial tasks. However, existing IKG construction methods still face core challenges such as data heterogeneity, complex semantic understanding, frequent knowledge changes, and limited automation. Inspired by the construction of IKG by industry experts, we propose CoMA-IKG, an large language model (LLM)-driven collaborative multiagent framework for automated construction of IKG. In the industrial data processing stage, an LLM-driven adaptive chunking agent is developed to achieve semantically complete and self-adjusting segmentation. In the triple extraction stage, a cluster of LLM-driven agents for progressive triple reasoning extraction and mechanism-aware logical discrimination is constructed to enable accurate industrial triple extraction under stepwise reasoning and industrial mechanism constraints. In the IKG evolution stage, an LLM-driven co-evolution agent is developed to generate evolution commands automatically based on the structural state of the IKG and real-time industrial data changes, enabling autonomous updating and continuous evolution of the IKG. Experimental results show that CoMA-IKG significantly outperforms existing automated knowledge graph construction methods in terms of relation mining, logical reasoning, and dynamic evolution of the IKG.

Grasping tasks are crucial in industrial manufacturing, where precise and efficient object grasping ensures smooth assembly processes and stable operations. Robotic arms, deployed in industrial environments, require timely and accurate computations to perform these tasks. This paper introduces BGRN, an RGB-D fusion-based binary grasp prediction network designed for lightweight grasp pose prediction in such settings. We propose a binary grasp pose prediction framework that significantly reduces memory usage by quantizing both weights and activations to 1 bit. Additionally, an interaction fusion module improves the integration of RGB and depth images, while an information recovery connection helps mitigate feature loss caused by binarization. Experimental results show that BGRN achieves competitive accuracy and notable reductions in memory usage and computational load compared to full-precision models.

Accurate demand forecasting for products in the parts supply chain system (PSCS) is critical for enterprises to optimize production and inventory operations. To address the challenges of inadequate modeling of complex inter-product relationships and the limited accuracy of existing forecasting methods, this paper proposes a demand forecasting method based on a Graph Convolutional and LSTM Network with Embedded Multi-Head Self-Attention (GL-MHSA). The method first extracts hybrid distance features, including Euclidean and pattern distances, from product sales data in the PSCS to mine product associations and construct graph-structured relational data. A Graph Convolutional Network (GCN) is then used to capture structural association features among products, while an LSTM network models the temporal dependencies in the demand sequences. The extracted features are fused through a Multi-Head Self-Attention (MHSA) mechanism to obtain a comprehensive feature representation. This representation is concatenated with other auxiliary features to form the final input for demand prediction. Experimental results on an automotive PSCS dataset show that the proposed GL-MHSA model achieves more accurate modeling of product associations and significantly improves demand forecasting performance compared to existing approaches.

In the context of Industrial Internet of Things (IIoT), time-series data is essential for maintenance and operational efficiency. However, challenges in IIoT data transmission, such as network instability, and in data annotation, like the high costs, lead to a shortage of high-quality labeled data, hindering system performance and industrial intelligence. Although traditional methods, such as signal imputation and denoising, have been employed, generative artificial intelligence (GAI) offers new possibilities for the generation of industrial time-series data. To address these challenges, we propose a novel score-based diffusion architecture specifically designed for industrial data generation. The score-based approach effectively leverages the gradients of the data distribution, offering a more structured and stable generative process compared to generative adversarial networks (GANs). Furthermore, our model incorporates predictive-corrective (PC) samplers with Langevin dynamics annealing to further optimize the generation process. Experimental results on turbofan engine datasets demonstrate that our model overcomes the inherent training instabilities of GANs, providing a more reliable and effective method for synthesizing high-fidelity industrial time-series data.

Recently, foundation models (such as ChatGPT) have emerged with powerful learning, understanding, and generalization abilities, showcasing tremendous potential to revolutionarily promote modern industry. Despite significant advancements in various fields, existing general foundation models face challenges in industry when dealing with the data of specialized modalities, the tasks of varying-scenario with multiple processes, and the requirements of trustworthy output, which makes industrial foundation model (IFM) a necessity. This article proposes a system architecture of termed IFMsys, including model training, model adaptation, and model application. Specifically, in model training, a base model is constructed by pretraining on multimodal industrial data and fine-tuning with fundamental industrial mechanisms. In model adaptation, the base model is developed into a series of task-oriented and domain-specific IFMs through fine-tuning with representative tasks and domain knowledge. In model application, an industrial agent-centric collaboration system and a comprehensive application framework of IFM are proposed to enhance the industrial product lifecycle applications. In addition, a prototype system of the IFM, namely, MetaIndux, is delivered, with application examples presented in typical industrial tasks. Finally, future research directions and open issues of IFM are prospected. We hope this article will inspire the advancements in the theories, technologies, and applications in this emerging research field of IFM.

Industrial time series prediction is highly important for the predictive maintenance of Industrial Internet of Things devices. Deep learning methods have demonstrated state-of-the-art (SOTA) performance in the field of time series prediction. However, time series data from complex industrial scenarios often contain substantial uncertainty. This makes it difficult for deterministic deep learning models to achieve accurate predictions. Moreover, existing static methods often fail to meet the real-time requirements of industrial environments. To address the challenges, this study introduces fuzzy learning into deep learning models to overcome the drawbacks of fixed model representations. Therefore, we propose a fuzzy dynamic transformer (FDformer) that can adaptively adjust network depth according to the complexity of individual samples. Subsequently, we design a fuzzy feature extraction mechanism to capture feature information within the fuzzy membership degree, enabling the feature-level fusion of the fuzzy representation with the dynamic depth representation. Finally, we propose a training method for dynamically allocating loss weights, emphasizing the contribution of various samples to different exits, thereby improving the performance of time-series dynamic networks. Experiments on multiple datasets indicate that FDformer achieves minimal computational costs and excellent prediction accuracy across multiple datasets, outperforming SOTA algorithms.

During the machining process of parts, choosing appropriate machining strategies optimizes production costs effectively. However, when the amount of data is limited, existing neural networks often struggle to fit the data accurately. Meanwhile, existing neural networks suffer from information dilution and lack effective mechanisms for direct information transfer between nonadjacent surfaces. This article proposes a method to extract General Machining Unit data. This data improves few-shot training performance. We investigate the distribution and information flow within General Machining Units and design a new way of data augmentation. In addition, to address the information dilution, a novel wormhole mechanism is proposed to aggregate information that spans the topological connections. In the backbone, we propose the Brep-WH layer that integrates wormhole mechanisms and attention pool layers. Both the Brep-WH network and the General Machining Unit data successfully improve the accuracy of the milling strategy dataset and the Fusion 360 Gallery segmentation dataset.

Complex industrial equipment plays a critical role in specific tasks within industrial edge scenarios. Predicting their health status accurately is essential to ensuring safety and reliability in the production process. However, real-world industrial edge scenarios often have limited resources and stringent real-time requirements, making it difficult to deploy high-precision deep learning models directly at the edge. To address this issue, this article proposes an efficient adaptive-gating binary neural network (ABNN). First, a trend-aware encoder (TAE) is proposed to optimize the binarization process of the input layer. Next, a learnable precision indicator (LPI) is proposed to adjust the inference precision level. Finally, an adaptive-gating convolution is proposed to improve the representational capabilities while maintaining the fitting ability without significantly increasing the computational cost. Additionally, a field-programmable gate array (FPGA) hardware accelerator is designed for the proposed network. ABNN achieves approximately a 7% improvement in accuracy and a 45% gain in efficiency compared to the baseline model.

Recently, deep learning-based models such as transformer have achieved significant performance for industrial remaining useful life (RUL) prediction due to their strong representation ability. In many industrial practices, RUL prediction algorithms are deployed on edge devices for real-time response. However, the high computational cost of deep learning models makes it difficult to meet the requirements of edge intelligence. In this article, a lightweight group transformer with multihierarchy time-series reduction (GT-MRNet) is proposed to alleviate this problem. Different from most existing RUL methods computing all time series, GT-MRNet can adaptively select necessary time steps to compute the RUL. First, a lightweight group transformer is constructed to extract features by employing group linear transformation with significantly fewer parameters. Then, a time-series reduction strategy is proposed to adaptively filter out unimportant time steps at each layer. Finally, a multihierarchy learning mechanism is developed to further stabilize the performance of time-series reduction. Extensive experimental results on the real-world condition datasets demonstrate that the proposed method can significantly reduce up to 74.7% parameters and 91.8% computation cost without sacrificing accuracy.

Implementing advanced AI techniques in industrial manufacturing requires large volumes of annotated sensor data. Unfortunately, collecting such data is often impractical due to extreme environments and the manual burden of expert annotation. Recent advancements in artificial intelligence generated content (AIGC) have inspired the exploration of industrial time-series generation to mitigate data shortages. However, existing AIGC models encounter difficulties in generating industrial time series due to their complex temporal dynamics, multichannel intercolumn correlations, and diverse frequency characteristics. To address these challenges, we propose MetaIndux-TS, a frequency-informed AIGC foundation model based on diffusion model frameworks. This model is designed to generate industrial time-series data under a variety of working conditions, across different types of equipment, and with variable lengths. Specifically, MetaIndux-TS integrates dual-frequency cross-attention networks, transforming time series into the frequency domain to model multivariate dependencies and capture intricate temporal details. In addition, the contrastive synthesis layer is constructed to generate high-fidelity time series by comparing periodic and long-term trends with initial noisy sequences. Comprehensive experiments show that MetaIndux-TS outperforms state-of-the-art models (SSSD, Dit, and TabDDPM), achieving a 57.5% improvement in fidelity and 20.4% in predictive score. MetaIndux-TS exhibits zero-shot generation capabilities for samples under unseen conditions, offering the potential to address data collection challenges in extreme environments. Codes are available at: https://github.com/Dolphin-wang/MetaIndux.

With the remarkable success of generative models like ChatGPT, artificial intelligence generated content (AIGC) is undergoing explosive development. Not limited to text and images, generative models can generate industrial time series data, addressing challenges, such as the difficulty of data collection and data annotation. Due to their outstanding generation ability, they have been widely used in Internet of Things, metaverse, and CPSS to enhance the efficiency of industrial production. In this article, we present a comprehensive overview of generative models for industrial time series from deep-generative models (DGMs) to large-generative models (LGMs). First, a DGM-based AIGC framework is proposed for industrial time series generation. Within this framework, we survey advanced industrial DGMs and present a multiperspective categorization. Then, we systematically propose the roadmap to construct industrial LGMs from four aspects: large-scale industrial dataset, LGMs architecture for complex industrial characteristics, self-supervised training for industrial time series, and fine-tuning of industrial downstream tasks. Furthermore, we introduce an evaluation benchmark that systematically assesses fidelity, diversity, and utility. We include a case study on aircraft engine maintenance, demonstrating the application of DGMs in industrial predictive maintenance. Finally, we conclude the challenges and future directions to enable the development of generative models in industry.

With the development of modern information technology and 5G, cloud-edge intelligent collaboration can make full use of the powerful computing power of cloud computing and the real-time response capability of edge computing to improve the overall efficiency of industrial intelligent systems. Thus, it shows strong application potential in digital twins, foundation models, and meta-universes. However, most of the existing researches focus on resource collaboration, data collaboration and application collaboration in cloud-edge computing framework, and there are many shortcomings in the research of intelligent collaboration framework. Therefore, we first propose a cloud-edge intelligent collaboration framework, which consists of three parts: terminal layer, edge artificial intelligence (AI) layer and cloud AI layer, including two core processes: cloud-edge intelligent training and cloud-edge intelligent inference. Then, we systematically analyze the key technologies of cloud-edge intelligent collaboration, including model segmentation, model early exit and foundation models. Finally, we put forward the application of cloud-edge intelligent collaboration in digital twin and AI Generated Content (AIGC), and through the analysis of typical application scenarios.

Invasive brain-computer interfaces with Electrocorticography (ECoG) have shown promise for high-performance speech decoding in medical applications, but less damaging methods like intracranial stereo-electroencephalography (sEEG) remain underexplored. With rapid advances in representation learning, leveraging abundant recordings to enhance speech decoding is increasingly attractive. However, popular methods often pre-train temporal models based on brain-level tokens, overlooking that brain activities in different regions are highly desynchronized during tasks. Alternatively, they pre-train spatial-temporal models based on channel-level tokens but fail to evaluate them on challenging tasks like speech decoding, which requires intricate processing in specific language-related areas. To address this issue, we collected a well-annotated Chinese word-reading sEEG dataset targeting language-related brain networks from 12 subjects. Using this benchmark, we developed the Du-IN model, which extracts contextual embeddings based on region-level tokens through discrete codex-guided mask modeling. Our model achieves state-of-the-art performance on the 61-word classification task, surpassing all baselines. Model comparisons and ablation studies reveal that our design choices, including (\romannumeral1) temporal modeling based on region-level tokens by utilizing 1D depthwise convolution to fuse channels in the ventral sensorimotor cortex (vSMC) and superior temporal gyrus (STG) and (\romannumeral2) self-supervision through discrete codex-guided mask modeling, significantly contribute to this performance. Overall, our approach -- inspired by neuroscience findings and capitalizing on region-level representations from specific brain regions -- is suitable for invasive brain modeling and represents a promising neuro-inspired AI approach in brain-computer interfaces. Code and dataset are available at https://github.com/liulab-repository/Du-IN.

Industrial multivariate time series (MTS) is a critical view of the industrial field for people to understand the state of machines. However, due to data collection difficulty and privacy concerns, available data for building industrial intelligence and industrial large models is far from sufficient. Therefore, industrial time series data generation is of great importance. Existing research usually applies generative adversarial networks (GANs) to generate MTS. However, GANs suffer from the unstable training process due to the joint training of the generator and discriminator. This article proposes a temporal-augmented conditional adaptive diffusion model, termed Diff-MTS, for MTS generation. It aims to better handle the complex temporal dependencies and dynamics of MTS data. Specifically, a conditional adaptive maximum-mean discrepancy (Ada-MMD) method has been proposed for the controlled generation of MTS, which does not require a classifier to control the generation. It improves the condition consistency of the diffusion model. Moreover, a temporal decomposition reconstruction UNet (TDR-UNet) is established to capture complex temporal patterns and further improve the quality of the synthetic time series. Comprehensive experiments on the C-MAPSS and FEMTO datasets demonstrate that the proposed Diff-MTS performs substantially better in terms of diversity, fidelity, and utility compared with the GAN-based methods. These results show that Diff-MTS facilitates the generation of industrial data, contributing to intelligent maintenance and the construction of industrial large models.

Prediction of industrial time-series is crucial for various Industrial Internet of Things applications. Despite the high accuracy of deep learning methods for time-series prediction, the significant memory requirements of deep learning models pose a challenge for the limited computational resources of industrial edge devices. To address this issue, this work proposes BTFormer, which achieves a high compression rate while maintaining competitive performance. First, a binary adaptive attention module is proposed to mitigate the loss of attention information caused by binarization. Second, a trend information soft-link is proposed to propagate trend information between layers and improve the representation ability of the model. Finally, a distribution-guided distillation strategy is proposed to optimize the training process. The experiments demonstrate that BTFormer effectively reduces model memory usage by 31.0 times and improves computational efficiency by 32.8 times while maintaining competitive performance.

Representation learning-based remaining useful life (RUL) prediction plays a crucial role in improving the security and reducing the maintenance cost of complex systems. Despite the superior performance, the high computational cost of deep networks hinders deploying the models on low-compute platforms. A significant reason for the high cost is the computation of representing long sequences. In contrast to most RUL prediction methods that learn features of the same sequence length, we consider that each time series has its characteristics and the sequence length should be adjusted adaptively. Our motivation is that an "easy" sample with representative characteristics can be correctly predicted even when short feature representation is provided, while "hard" samples need complete feature representation. Therefore, we focus on sequence length and propose a dynamic length transformer (DLformer) that can adaptively learn sequence representation of different lengths. Then, a feature reuse mechanism is developed to utilize previously learned features to reduce redundant computation. Finally, in order to achieve dynamic feature representation, a particular confidence strategy is designed to calculate the confidence level for the prediction results. Regarding interpretability, the dynamic architecture can help human understand which part of the model is activated. Experiments on multiple datasets show that DLformer can increase up to 90% inference speed, with less than 5% degradation in model accuracy.